Skip to main content
Back to blog
•2 min read

DeepSeek Changed Everything Overnight

AIdeepseekindustry

okay so something wild happened.

a chinese company called deepseek released deepseek-r1, and within days it was the most downloaded app on the iOS app store. ahead of chatgpt.

the AI world is... processing.

what happened

deepseek dropped deepseek-v3 and then deepseek-r1 in quick succession. r1 is specifically tuned for reasoning—think chain-of-thought, complex problem solving.

the benchmarks are strong. really strong. competitive with gpt-4 and claude on many tasks.

but here's the thing that has everyone talking: they did it for a fraction of the cost.

the cost story

the claims (not fully verified but widely discussed):

  • trained on fewer gpus than assumed possible
  • much lower training costs than western labs
  • still achieved competitive results

if true, this challenges a lot of assumptions about what it takes to build frontier models.

my reaction as someone at anthropic

honestly? mixed.

the positive spin: competition is good. more capable models from more places means faster progress for humanity.

the concerning spin: the safety considerations in chinese AI development are... different. regulatory environment, transparency, access—all different.

the pragmatic spin: this was inevitable. AI capability was never going to stay concentrated in a few us companies.

trying the model

i tried deepseek-r1 (they have a free api).

it's good. genuinely impressive for coding and reasoning tasks. the chain-of-thought is interesting to watch.

is it "better" than claude? depends on the task. but it's clearly in the same tier.

what this means

1. the cost curve matters if you can train frontier models for less, more players can enter the game.

2. open source is leveling up deepseek released weights. this helps the entire ecosystem.

3. china is serious anyone who thought chinese AI was years behind needs to revise that estimate.

4. regulatory questions intensify how do we think about safety and oversight when models come from different governance systems?

the bigger conversation

i don't have answers here. the geopolitics of AI is complicated and above my pay grade.

what i do know: the world just got more multipolar in terms of AI capability. that has implications.

watching this space closely. the pace of change is accelerating in ways nobody predicted.